Science Watch
Concepts and Conceptual Structure
Douglas L. Medin University of Illinois at Urbana
ABSTRACT: Research and theory on categorization and
conceptual structure have recently undergone two major
shifts. The first shift is from the assumption that concepts
have defining properties (the classical view) to the idea
that concept representations may be based on properties
that are only characteristic or typical of category examples
(the probabilistic view). Both the probabilistic view and
the classical view assume that categorization is driven by
similarity relations. A major problem with describing cat-
egory structure in terms of similarity is that the notion of
similarity is too unconstrained to give an account of con-
ceptual coherence. The second major shift is from the idea
that concepts are organized by similarity to the idea that
concepts are organized around theories. In this article,
the evidence and rationale associated with these shifts are
described, and one means of integrating similarity-based
and theory-driven categorization is outlined.
What good are categories? Categorization involves treat-
ing two or more distinct entities as in some way equivalent
in the service of accessing knowledge and making pre-
dictions. Take psychodiagnostic categories as an example.
The need to access relevant knowledge explains why clin-
ical psychologists do not (or could not) treat each indi-
vidual as unique. Although one would expect treatment
plans to be tailored to the needs of individuals, absolute
uniqueness imposes the prohibitive cost of ignorance.
Clinicians need some way to bring their knowledge and
experience to bear on the problem under consideration,
and that requires the appreciation of some similarity or
relationship between the current situation and what has
gone before. Although clinical psychologists may or may
not use a specific categorization system, they must find
points of contact between previous situations and the
current context; that is, they must categorize. Diagnostic
categories allow clinicians to predict the efficacy of alter-
native treatments and to share their experiences with other
therapists. Yet another reason to categorize is to learn
about etiology. People who show a common manifestation
of some problem may share common precipitating con-
ditions or causes. Ironically, the only case in which cat-
egorization would not be useful is where all individuals
are treated alike; thus, categorization allows diversity.
More generally speaking, concepts and categories
serve as building blocks for human thought and behavior.
Roughly, a concept is an idea that includes all that is char-
acteristically associated with it. A category is a partitioning
or class to which some assertion or set of assertions might
apply. It is tempting to think of categories as existing in
the world and of concepts as corresponding to mental
representations of them, but this analysis is misleading.
It is misleading because concepts need not have real-world
counterparts (e.g., unicorns) and because people may im-
pose rather than discover structure in the world. I believe
that questions about the nature of categories may be psy-
chological questions as much as metaphysical questions.
Indeed, for at least the last decade my colleagues and I
have been trying to address the question of why we have
the categories we have and not others. The world could
be partitioned in a limitless variety of ways, yet people
find only a miniscule subset of possible classifications to
be meaningful. Part of the answer to the categorization
question likely does depend on the nature of the world,
but part also surely depends on the nature of the organism
and its goals. Dolphins have no use for psychodiagnostic
categories.
Given the fundamental character of concepts and
categories, one might think that people who study con-
cepts would have converged on a stable consensus with
respect to conceptual structure. After all, Plato and Ar-
istotle had quite a bit to say about concepts, medieval
philosophers were obsessed with questions about univer-
sals and the essence of concepts, and concept represen-
tation remains as a cornerstone issue in all aspects of
cognitive science. However, we have neither consensus
nor stability. The relatively recent past has experienced
at least one and probably two major shifts in thought
about conceptual structure, and stability is the least salient
attribute of the current situation. In the remainder of this
article, I will briefly describe these shifts and then outline
some ways of integrating the strong points of the various
views.
The First Shift: Classical Versus
Probabilistic Views
It is difficult to discuss concepts without bringing in the
notion of similarity at some point. For example, a com-
mon idea is that our classification system tends to max-
imize within-category similarity relative to between-cat-
egory similarity. That is, we group things into categories
because they are similar. It will be suggested that alter-
December 1989
• American Psychologist
Copyright 1989 by the American Psychological Association, Inc. 0003-066X/89/$00.75
Vol. 44, No. 12, 1469-1481
1469
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
native views of conceptual structure are associated with
distinct (though sometimes implicit) theories of the nature
of similarity.
The Classical View
The idea that all instances or examples of a category have
some fundamental characteristics in common that de-
termine their membership is very compelling. The clas-
sical view of concepts is organized around this notion.
The classical view assumes that mental representations
of categories consist of summary lists of features or prop-
erties that individually are necessary for category mem-
bership and collectively are sufficient to determine cate-
gory membership. The category triangle meets these cri-
teria. All triangles are closed geometric forms with three
sides and interior angles that sum to 180 degrees. To see
if something is a triangle one has only to check for these
three properties, and if any one is missing one does not
have a triangle.
What about other concepts? The classical view sug-
gests that all categories have defining features. A particular
person may not know what these defining features are
but an expert certainly should. In our 1981 book, Cate-
gories and Concepts, Ed Smith and I reviewed the status
of the classical view as a theory of conceptual structure.
We concluded that the classical view was in grave trouble
for a variety of reasons. Many of the arguments and
counterarguments are quite detailed, but the most serious
problems can be easily summarized:
1. Failure to specify defining features. One glaring
problem is that even experts cannot come up with defin-
ing features for most lexical concepts (i.e., those reflected
in our language). People may believe that concepts have
necessary or sufficient features (McNamara & Sternberg,
1983), but the features given as candidates do not hold
up to closer scrutiny. For example, a person may list
"made of wood" as a necessary property for violins, but
not all violins are made of wood. Linguists, philosophers,
biologists, and clinical psychologists alike have been un-
able to supply a core set of features that all examples of
a concept (in their area of expertise) necessarily must
share.
2, Goodness of example effects. According to the
classical view, all examples of a concept are equally good
because they all possess the requisite defining features.
Experience and (by now) a considerable body of research
undermines this claim. For example, people judge a robin
to be a better example of bird than an ostrich is and can
answer category membership questions more quickly for
good examples than for poor examples (Smith, Shoben,
& Rips, 1974). Typicality effects are nearly ubiquitous
The research described in this article was supported in part by National
Science Foundation Grant No. BNS 84-19756 and by National Library
of Medicine Grant No. LM 04375. Brian Ross, Edward Shoben, Ellen
Markman, Greg Oden, and Dedre Gentner provided helpful comments
on an earlier draft of the article.
Correspondence concerning this article should be addressed to
Douglas L. Medin, who is now at Department of Psychology, University
of Michigan, 330 Packard Road, Ann Arbor, MI 48104.
(for reviews, see Medin & Smith, 1984; Mervis & Rosch,
1981; Oden, 1987); they hold for the artistic style (Hartley
& Homa, 1981), chess (Goldin, 1978), emotion terms
(Fehr, 1988; Fehr & Russell, 1984), medical diagnosis
(Arkes & Harkness, 1980), and person perception (e.g.,
Cantor & Mischel, 1977).
Typicality effects are not, in principle, fatal for the
classical view. One might imagine that some signs or fea-
tures help to determine the presence of other (defining)
features. Some examples may have more signs or clearer
signs pointing the way to the defining properties, and this
might account for the difference in goodness of example
judgments or response times. This distinction between
identification procedures (how one identifies an instance
of a concept) and a conceptual core (how the concept
relates to other concepts) may prove useful if it can be
shown that the core is used in some other aspect of think-
ing. It seems, however, that this distinction serves more
to insulate the classical view from empirical findings, and
Smith, Rips, and Medin (1984) argued that there are no
sharp boundaries between core properties and those used
for purposes of identification.
3. Unclear cases. The classical view implies a pro-
cedure for unambiguously determining category mem-
bership; that is, check for defining features. Yet there are
numerous cases in which it is not clear whether an ex-
ample belongs to a category. Should a rug be considered
furniture? What about a clock or radio? People not only
disagree with each other concerning category membership
but also contradict themselves when asked about mem-
bership on separate occasions (Barsalou, 1989; Bellezza,
1984; McCloskey & Glucksberg, 1978).
These and other problems have led to disenchant-
ment with the classical view of concepts. The scholarly
consensus has shifted its allegiance to an alternative, the
probabilistic view.
The Probabilistic View
The rejection of the classical view of categories has been
associated with the ascendance of the probabilistic view
of category structure (Wittgenstein, 1953). This view
holds that categories are "fuzzy" or ill-defined and that
categories are organized around a set of properties or
clusters of correlated attributes (Rosch, 1975) that are
only characteristic or typical of category membership.
Thus, the probabilistic view rejects the notion of defining
features.
The most recent edition of the Diagnostic and Sta-
tistical Manual of Mental Disorders (DSM-IIIR, Amer-
ican Psychiatric Association, 1987) uses criteria based on
lists of characteristic symptoms or features to describe
diagnostic categories and thereby endorses the probabi-
listic view. For example, a diagnosis of depression can be
made if a dysphoric mood and any five of a set of nine
symptoms are present nearly every day for a period of at
least two weeks. Thus, two people may both be categorized
as depressed and share only a single one of the nine char-
acteristic symptoms[
The probabilistic view is perfectly at home with the
1470 December 1989 • American Psychologist
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
typicality effects that were so awkward for the classical
view. Membership in probabilistic categories is naturally
graded, rather than all or none, and the better or more
typical members have more characteristic properties than
the poorer ones. It is also easy to see that the probabilistic
view may lead to unclear cases. Any one example may
have several typical properties of a category but not so
many that it clearly qualifies for category membership.
In some pioneering work aimed at clarifying the
structural basis of fuzzy categories, Rosch and Mervis
(1975) had subjects list properties of exemplars for a va-
riety of concepts such as bird, fruit, and tool. They found
that the listed properties for some exemplars occurred
frequently in other category members, whereas others had
properties that occurred less frequently. Most important,
the more frequently an excmplar's properties appeared
within a category, the higher was its rated typicality for
that category. The correlation between number of char-
acteristic properties possessed and typicality rating was
very high and positive. For example, robins have char-
acteristic bird properties of flying, singing, eating worms,
and building nests in trees, and they are rated to be very
typical birds. Penguins have none of these properties, and
they are'rated as very atypical birds. In short, the Rosch
and Mervis work relating typicality to number of char-
acteristic properties put the probabilistic view on fairly
firm footing.
1. Mental representations of probabilistic view cat-
egories. If categories arc not represented in terms of
definitions, what form do our mental representations
take? The term, probabilistic view, seems to imply that
people organize categories via statistical reasoning. Ac-
tually, however, there is a more natural interpretation of
fuzzy categories. Intuitively, probabilistic view categories
are organized according to a family resemblance principle.
A simple form of summary representation would be an
example or ideal that possessed all of the characteristic
features of a category. This summary representation is
referred to as the prototype, and the prototype can be
used to decide category membership. If some candidate
example is similar enough to the prototype for a category,
then it will be classified as a member of that category.
The general notion is that, based on experience with ex-
amples of a category, people abstract out the central ten-
dency or prototype that becomes the summary mental
representation for the category.
A more radical principle of mental representation,
which is also consistent with fuzzy categories, is the ex-
emplar view (Smith & Medin, 1981). The exemplar view
denies that there is a single summary representation and
instead claims that categories are represented by means
of examples. In this view, clients may be diagnosed as
suicidal, not because they are similar to some prototype
of a suicidal person, but because they remind the clinician
of a previous client who was suicidal.
A considerable amount of research effort has been
aimed at contrasting exemplar and prototype represen-
tations (see Allen, Brooks, Norman, & Rosenthal, 1988;
Estes, 1986a, 1986b; Medin, 1986; Medin & Smith, 1984;
Nosofsky, 1987, 1988a; and Odcn, 1987). Genero and
Cantor (1987) suggested that prototypes serve untrained
diagnosticians well but that trained diagnosticians may
find exemplars to be more helpful. For my present pur-
poses, however, I will blur over this distinction to note
that both prototype and exemplar theories rely on roughly
the same similarity principle. That is, category member-
ship is determined by whether some candidate is suffi-
ciently similar either to the prototype or to a set of en-
coded examples, where similarity is based on matches
and mismatches of independent, equally abstract, fea-
tures.
2. Probabilistic view and similarity. To give mean-
ing to thc claim that categorization is based on similarity,
it is important to be specific about what one means by
similarity. Although the consensus is not uniform, I be-
lieve that the modal model of similarity with respect to
conceptual structure can be summarized in terms of the
four assumptions as follows: (a) Similarity between two
things increases as a function of the number of features
or properties they share and decreases as a function of
mismatching or distinctive features. (b) These features
can be treated as independent and additive. (c) The fea-
tures determining similarity are all roughly the same level
of abstractness (as a special case they may be irreducible
primitives). (d) These similarity principles are sufficient
to describe conceptual structure, and therefore, a concept
is more or less equivalent to a list of its features. This
theory of similarity is very compatible with the notion
that categories arc organized around prototypes. None-
theless, I will later argue that each of these assumptions
is wrong or misleading and that to understand conceptual
structure theories of similarity are needed that reject each
of these assumptions. Before outlining an alternative set
of similarity assumptions, however, I will first describe a
set of observations that motivate the second, still more
recent, shift in thinking concerning conceptual structure.
Problems for Probabilistic View Theories
Problems for Prototypes
Although the general idea that concepts are organized
around prototypes remains popular, at a more specific,
empirical level, prototype theories have not fared very
well. First of all, prototype theories treat concepts as con-
text-independent. Roth and Shoben (1983), however, have
shown that typicality judgments vary as a function of
particular contexts. For example, tea is judged to be a
more typical beverage than milk in the context of sec-
retaries taking a break, but this ordering reverses for the
context of truck drivers taking a break. Similarly, Shoben
and I (Medin & Shoben, 1988) noted that the typicality
of combined concepts cannot be predicted from the typ-
icality of the constituents. As an illustrative example,
consider the concept of spoon. People rate small spoons
as more typical spoons than large spoons, and metal
spoons as more typical spoons than wooden spoons. If
the concept spoon is represented by a prototypic spoon,
then a small metal spoon should be the most typical
December 1989
•
American
Psychologist
1471
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
spoon, followed by small wooden and large metal spoons,
and large wooden spoons should be the least typical. In-
stead, people find large wooden spoons to be more typical
spoons than either small wooden spoons or large metal
spoons (see also Malt & Smith, 1983). The only way for
a prototype model to handle these results is to posit mul-
tiple prototypes. But this strategy creates new problems.
Obviously one cannot have a separate prototype for every
adjective noun combination because there are simply too
many possible combinations. One might suggest that there
are distinct subtypes for concepts like
spoon,
but one
would need a theory describing how and when subtypes
are created. Current prototype models do not provide
such a theory. A third problem for prototype theories
grows out of Barsalou's work (1985, 1987) on goal-derived
categories such as "things to take on a camping trip" and
"foods to eat while on a diet." Barsalou has found that
goal-derived categories show the same typicality effects
as other categories. The basis for these effects, however,
is not similarity to an average or prototype but rather
similarity to an ideal. For example, for the category of
things to eat while on a diet, typicality ratings are deter-
mined by how closely an example conforms to the ideal
of zero calories.
Laboratory studies of categorization using artificially
constructed categories also raise problems for prototypes.
Normally many variables relevant to human classification
are correlated and therefore confounded with one another.
The general rationale for laboratory studies with artifi-
cially created categories is that one can isolate some vari-
able or set of variables of interest and unconfound some
natural correlations. Salient phenomena associated with
fuzzy categories are observed with artificially constructed
categories, and several of these are consistent with pro-
totype theories. For example, one observes typicality ef-
fects in learning and on transfer tests using both correct-
ness and reaction time as the dependent variable (e.g.,
Rosch & Mervis, 1975). A striking phenomenon, readily
obtained, is that the prototype for a category may be clas-
sified more accurately during transfer tests than are the
previously seen examples that were used during original
category learning (e.g., Homa & Vosburgh, 1976; Medin
& Schaffer, 1978; Peterson, Meagher, Chait, & Gillie,
1973).
Typicality effects and excellent classification of pro-
totypes are consistent with the idea that people are learn-
ing these ill-defined categories by forming prototypes.
More detailed analyses, however, are more problematic.
Prototype theory implies that the only information ab-
stracted from categories is the central tendency. A pro-
totype representation discards information concerning
category size, the variability of the examples, and infor-
mation concerning correlations of attributes. The evi-
dence suggests that people are sensitive to all three of
these types of information (Estes, 1986b; Flannagan,
Fried, & Holyoak, 1986; Fried & Holyoak, 1984; Medin,
Altom, Edelson, & Freko, 1982; Medin & Schaffer, 1978).
An example involving correlated attributes pinpoints part
of the problem. Most people have the intuition that small
birds are much more likely to sing than large birds. This
intuition cannot be obtained from a single summary pro-
totype for birds. The fact that one can generate large
numbers of such correlations is a problem for the idea
that people reason using prototypes. More generally, pro-
totype representations seem to discard too much infor-
mation that can be shown to be relevant to human cat-
egorizations.
Yet another problem for prototypes is that they make
the wrong predictions about which category structures
should be easy or difficult to learn. One way to concep-
tualize the process of classifying examples on the basis
of similarity to prototypes is that it involves a summing
of evidence against a criterion. For example, if an instance
shows a criterial sum of features (appropriately weighted),
then it will be classified as a bird, and the more typical a
member is of the category, the more quickly the criterion
will be exceeded. The key aspect of this prediction is that
there must exist some additive combination of properties
and their weights that can be used to correctly assign
instances as members or nonmembers. The technical
term for this constraint is that categories must be linearly
separable (Sebestyn, 1962). For a prototype process to
work in the sense of accepting all members and rejecting
all nonmembers, the categories must be linearly separable.
If linear separability acts as a constraint on human
categorization, then with other factors equal, people
should find it easier to learn categories that are linearly
separable than categories that are not linearly separable.
To make a long story short, however, studies employing
a variety of stimulus materials, category sizes, subject
populations, and instructions have failed to find any ev-
idence that linear separability acts as a constraint on hu-
man classification learning (Kemler-Nelson, 1984; Medin
& Schwanenflugel, 1981; see also Shepard, Hovland, &
Jenkins, 1961).
The cumulative effect of these various chunks of ev-
idence has been to raise serious questions concerning the
viability of prototype theories. Prototype theories imply
constraints that are not observed in human categoriza-
tion, predict insensitivity to information that people
readily use, and fail to reflect the context sensitivity that
is evident in human categorization. Rather than getting
at the character of human conceptual representation,
prototypes appear to be more of a caricature of it. Ex-
emplar models handle some of these phenomena, but they
fail to address some of the most fundamental questions
concerning conceptual structure.
Exemplar-Based
Theories
The problems just described hold not only for prototype
theories in particular but also for any similarity-based
categorization model that assumes that the constituent
features are independent and additive. To give but one
example, one could have an exemplar model of catego-
rization that assumes that, during learning, people store
examples but that new examples are classified by "com-
puting" prototypes and determining the similarity of the
novel example to the newly constructed prototypes. In
1472 December 1989 • American Psychologist
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
short, the central tendency would be abstracted (and other
information discarded) at the time of retrieval rather than
at the time of storage or initial encoding. Such a model
would inherit all the shortcomings of standard prototype
theories.
Some exemplar storage theories do not endorse the
notion of feature independence (Hintzman, 1986; Medin
& Schaffer, 1978), or they assume that classification is
based on retrieving only a subset of the stored examples
(presumably the most similar ones or, as a special case,
the most similar one). The idea that retrieval is limited,
similarity-based, and context-sensitive is in accord with
much of the memory literature (e.g., Tulving, 1983). In
addition, these exemplar models predict sensitivity to
category size, instance variability, context, and correlated
attributes. It is my impression that in head-to-head com-
petition, exemplar models have been substantially more
successful than prototype models (Barsalou & Medin,
1986; Estes, 1986b; Medin & Ross, 1989; Nosofsky,
1988a, 1988b; but see Homa, 1984, for a different
opinion).
Why should exemplar models fare better than pro-
totype models? One of the main functions of classification
is that it allows one to make inferences and predictions
on the basis of partial information (see Anderson, 1988).
Here I am using classification loosely to refer to any means
by which prior (relevant) knowledge is brought to bear,
ranging from a formal classification scheme to an idio-
syncratic reminding of a previous case (which, of course,
is in the spirit of exemplar models; see also Kolodner,
1984). In psychotherapy, clinicians are constantly making
predictions about the likelihood of future behaviors or
the efficacy of a particular treatment based on classifi-
cation. Relative to prototype models, exemplar models
tend to be conservative about discarding information that
facilitates predictions. For instance, sensitivity to corre-
lations of properties within a category enables finer pre-
dictions: From noting that a bird is large, one can predict
that it cannot sing. It may seem that exemplar models
do not discard any information at all, but they are in-
complete without assumptions concerning retrieval or
access. In general, however, the pairs of storage and re-
trieval assumptions associated with exemplar models
preserve much more information than prototype models.
In a general review of research on categorization and
problem-solving, Brian Ross and I concluded that ab-
straction is both conservative and tied to the details of
specific examples in a manner more in the spirit of ex-
emplar models than prototype models (Medin & Ross,
1989).
Unfortunately, context-sensitive, conservative cate-
gorization is not enough. The debate between prototype
and exemplar models has taken place on a platform con-
structed in terms of similarity-based categorization. The
second shift is that this platform has started to crumble,
and the viability of probabilistic view theories of cate-
gorization is being seriously questioned. There are two
central problems. One is that probabilistic view theories
do not say anything about why we have the categories we
have. This problem is most glaringly obvious for exemplar
models that appear to allow any set of examples to form
a category. The second central problem is with the notion
of similarity. Do things belong in the same category be-
cause they are similar, or do they seem similar because
they are in the same category?
Does
Similarity Explain Categorization?
1. Flexibility. Similarity is a very intuitive notion. Un-
fortunately, it is even more elusive than it is intuitive.
One problem with using similarity to define categories is
that similarity is too flexible. Consider, for example,
Tversky's (1977) influential contrast model, which defines
similarity as a function of common and distinctive fea-
tures weighted for salience or importance. According to
this model, similarity relationships will depend heavily
on the particular weights given to individual properties
or features. For example, a zebra and a barberpole would
be more similar than a zebra and a horse if the feature
"striped" had sufficient weight. This would not necessarily
be a problem if the weights were stable. However, Tversky
and others have convincingly shown that the relative
weighting of a feature (as well as the relative importance
of matching and mismatching features) varies with the
stimulus context, experimental task (Gati & Tversky,
1984; Tversky, 1977), and probably even the concept un-
der consideration (Ortony, Vondruska, Foss, & Jones,
1985). For example, common properties shared by a pair
of entities may become salient only in the context of some
third entity that does not share these properties.
Once one concedes that similarity is dynamic and
depends on some (not well-understood) processing prin-
ciples, earlier work on the structural underpinnings of
fuzzy categories can be seen in a somewhat different light.
Recall that the Rosch and Mervis (1975) studies asked
subjects to list attributes or properties of examples and
categories. It would be a mistake to assume that people
had the ability to read and report their mental represen-
tations of concepts in a veridical manner. Indeed Keil
(1979, 1981) pointed out that examples like robin and
squirrel shared many important properties that almost
never show up in attribute listings (e.g., has a heart,
breathes, sleeps, is an organism, is an object with bound-
aries, is a physical object, is a thing, can be thought about,
and so on). In fact, Keil argued that knowledge about
just these sorts of predicates, referred to as ontological
knowledge (Sommers, 1971), serves to organize children's
conceptual and semantic development. For present pur-
poses, the point is that attribute listings provide a biased
sample of people's conceptual knowledge. To take things
a step further, one could argue that without constraints
on what is to count as a feature, any two things may be
arbitrarily similar or dissimilar. Thus, as Murphy and I
(Murphy & Medin, 1985) suggested, the number of prop-
erties that plums and lawn mowers have in common could
be infinite: Both weigh less than 1000 Kg, both are found
on earth, both are found in our solar system, both cannot
hear well, both have an odor, both are not worn by ele-
phants, both are used by people, both can be dropped,
December 1989 • American Psychologist 1473
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
and so on (see also Goodman, 1972; Watanabe, 1969).
Now consider again the status of attribute listings. They
represent a biased subset of stored or readily inferred
knowledge. The correlation of attribute listings with typ-
icality judgments is a product of such knowledge and a
variety of processes that operate on it. Without a theory
of that knowledge and those processes, it simply is not
clear what these correlations indicate about mental rep-
resentations.
The general point is that attempts to describe cat-
egory structure in terms of similarity will prove useful
only to the extent that one specifies which principles de-
termine what is to count as a relevant property and which
principles determine the importance of particular prop-
erties. It is important to realize that the explanatory work
is being done by the principles which specify these con-
straints rather than the general notion of similarity. In
that sense similarity is more like a dependent variable
than an independent variable.
2. Attribute matching and categorization.
The
modal model of similarity summarized in Table 1 invites
one to view categorization as attribute matching. Al-
though that may be part of the story, there are several
ways in which the focus on attribute matching may be
misleading. First of all, as Armstrong, Gleitman, and
Gleitman (1983) emphasized, most concepts are not a
simple sum of independent features. The features that
are characteristically associated with the concept
bird
are
just a pile of bird features unless they are held together
in a "bird structure." Structure requires both attributes
and
relations
binding the attributes together. Typical bird
features (laying eggs, flying, having wings and feathers,
building nests in trees, and singing) have both an internal
structure and an external structure based on interproperty
relationships. Building nests is linked to laying eggs, and
building nests in trees poses logistical problems whose
solution involves other properties such as having wings,
flying, and singing. Thus, it makes sense to ask why birds
have certain features (e.g., wings and feathers). Although
people may not have thought about various interproperty
relationships, they can readily reason with them. Thus,
one can answer the question of why birds have wings and
feathers (i.e., to fly).
In a number of contexts, categorization may be more
like problem solving than attribute matching. Inferences
and causal attributions may drive the categorization pro-
cess. Borrowing again from work by Murphy and me
(1985), "jumping into a swimming pool with one's clothes
on" in all probability is not associated directly with the
concept
intoxicated.
However, observing this behavior
might lead one to classify the person as drunk. In general,
real world knowledge is used to reason about or explain
properties, not simply to match them. For example, a
teenage boy might show many of the behaviors associated
with an eating disorder, but the further knowledge that
the teenager is on the wrestling team and trying to make
a lower weight class may undermine any diagnosis of a
disorder.
3. Summary.
It does not appear that similarity, at
least in the form it takes in current theories, is going
to
be at all adequate to explain categorization. Similarity
may be a byproduct of conceptual coherence rather than
a cause. To use a rough analogy, winning basketball teams
have in common scoring more points than their oppo-
nents, but one must turn to more basic principles to ex-
plain why they score more points. One candidate for a
set of deeper principles is the idea that concepts are or-
ganized around theories, and theories provide conceptual
coherence. In the next section, I will briefly summarize
some of the current work on the role of knowledge struc-
tures and theories in categorization and then turn to a
form of rapprochement between similarity and knowl-
edge-based categorization principle.
The Second Shift: Concepts as Organized
by Theories
Knowledge-Based Categorization
It is perhaps only a modest exaggeration to say that sim-
ilarity gets at the shadow rather than the substance of
concepts. Something is needed to give concepts life, co-
herence, and meaning. Although many philosophers of
science have argued that observations are necessarily the-
ory-labeled, only recently have researchers begun to stress
that the organization of concepts is knowledge-based and
driven by theories about the world (e.g., Carey, 1985; S.
Gelman, 1988; S. Gelman & Markman, 1986a, 1986b;
Keil, 1986, 1987; Keil & Kelly, 1987; Lakoff, 1987;
Markman, 1987; Massey & R. Gelman, 1988; Murphy
& Medin, 1985; Oden, 1987; Rips, 1989; Schank, Collins,
& Hunter, 1986; and others).
The primary differences between the similarity-
based and theory-based approaches to categorization are
summarized in Table 1, taken from Murphy and Medin
(1985). Murphy and Medin suggested that the relation
between a concept and an example is analogous to
the
relation between theory and data. That is, classification
is not simply based on a direct matching of properties of
the concept with those in the example, but rather requires
that the example have the right "explanatory relation-
ship" to the theory organizing the concept. In the case of
a person diving into a swimming pool with his or her
clothes on, one might try to reason back to either causes
or predisposing conditions. One might believe that having
too
much to drink impairs judgment and that going into
the pool shows poor judgment. Of course, the presence
of other information, such as the fact that another person
who cannot swim has fallen into the pool, would radically
change the inferences drawn and, as a consequence,
the
categorization judgment.
One of the more promising aspects of the theory-
based approach is that it begins to address the question
of why we have the categories we have or why categories
are sensible. In fact, coherence may be achieved in the
absence of any obvious source of similarity among ex-
amples. Consider the category comprised of children,
money, photo albums, and pets. Out of context the cat-
egory seems odd. If one's knowledge base is enriched to
1474 December 1989 • American Psychologist
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
I II I
Table 1
Comparison of Two Approaches to Concepts
III
Aspect of
conceptual theory Similarity-based approach Theory-based approach
Concept representation
Category definition
Units of analysis
Similarity
structure, attribute lists,
correlated attributes
Various similarity metrics,
summation of attributes
Attributes
Categorization basis
Attribute matching
Weighting of attributes
Cue validity, salience
Interconceptual
structure
Hierarchy
based on shared
attributes
Conceptual Feature
accretion
development
I
include the fact that the category represents "things to
take out of one's house in case of a fire," the category
becomes sensible (Barsalou, 1983). In addition, one could
readily make judgments about whether new examples
(e.g., personal papers) belonged to the category, judgments
that would not be similarity based.
Similarity effects can be overridden by theory-related
strategies even in the judgments of young children. That
fact was very nicely demonstrated by Gelman and Mark-
man (1986a) in their studies of induction. Specifically,
they pitted category membership against perceptual sim-
ilarity in an inductive inference task. Young children were
taught that different novel properties were true of two
examples and then were asked which property was also
true of a new example that was similar to one alternative
but belonged to a different category, and one that was
perceptually different from the other examples but be-
longed to the same category. For example, children might
be taught that a (pictured) flamingo feeds its baby mashed-
up food and that a (pictured) bat feeds its baby milk, and
then they might be asked how a (pictured) owl feeds its
baby. The owl was more perceptually similar to the bat
than to the flamingo, but even four-year-olds made in-
ferences on the basis of category membership rather than
similarity.
Related work by Susan Carey and Frank Keil shows
that children's biological theories guide their conceptual
development. For example, Keil has used the ingenious
technique of describing transformations or changes such
as painting a horse to look like a zebra to examine the
extent to which category membership judgments are
controlled by superficial perceptual properties. Biological
theories determine membership judgments quite early on
(Keil, 1987; Keil & Kelly, 1987). Rips (1989) has used
the same technique to show that similarity is neither nee-
Correlated attributes plus underlying principles that
determine
which correlations are noticed
An explanatory principle common to category
members
Attributes plus explicitly represented relations
of
attributes and concepts
Matching plus inferential processes supplied
by
underlying principles
Determined in part by importance in the underlying
principles
Network formed by causal and explanatory links, as
well as sharing of properties picked out as
relevant
Changing organization and explanations
of
concepts as a result of world knowledge
II IIII
essary nor sufficient to determine category membership.
It even appears to be the case that theories can affect
judgments of similarity. For example, Medin and Shoben
(1988) found that the terms
white hair
and
grey hair
were
judged to be more similar than grey
hair and black hair,
but that the terms
white clouds
and
grey clouds
were
judged as less similar than
grey clouds and black clouds.
Our interpretation is that white and grey hair are linked
by a theory (of aging) in a way that white and grey clouds
are not.
The above observations are challenging for defenders
of the idea that similarity drives conceptual organization.
In fact, one might wonder if the notion of similarity is so
loose and unconstrained that we might be better offwith-
out it. Goodman (1972) epitomized this attitude by calling
similarity "a pretender, an imposter, a quack" (p. 437).
After reviewing some reasons to continue to take simi-
larity seriously, I outline one possible route for integrating
similarity-based and theory-based categorization.
The
Need for Similarity
So far I have suggested that similarity relations do not
provide conceptual coherence but that theories do. Be-
cause a major problem with similarity is that it is so un-
constrained, one might ask what constrains theories. If
we cannot identify constraints on theories, that is, say
something about why we have the theories we have and
not others, then we have not solved the problem of co-
herence: It simply has been shifted to another level. Al-
though I believe we can specify some general properties
of theories and develop a psychology of explanation (e.g.,
Abelson & Lalljee, 1988; Einhorn & Hogarth, 1986; Hil-
ton & Slugoski, 1986; Leddo, Abelson, & Gross, 1984),
I equally believe that a constrained form ofslmilarity will
play an important role in our understanding of human
December 1989 • American Psychologist 1475
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
concepts. This role is not to provide structure so much
as it is to guide learners toward structure.
The impact of more direct perceptual similarity on
the development of causal explanations is evident in the
structure of people's naive theories. Frazer's (1959) cross-
cultural analysis of belief systems pointed to the ubiquity
of two principles, homeopathy and contagion. The prin-
ciple of homeopathy is that causes and effects tend to be
similar. One manifestation of this principle is homeo-
pathic medicine, in which the cure (and the cause) are
seen to resemble the symptoms. In the Azande culture,
for example, the cure for ringworm is to apply fowl's
excrement because the excrement looks like the ring-
worm. Schweder (1977) adduced strong support for the
claim that resemblance is a fundamental conceptual tool
of everyday thinking in all cultures, not just so-called
primitive cultures.
Contagion is the principle that a cause must have
some form of contact to transmit its effect. In general,
the more contiguous (temporally and spatially similar)
events are in time and space, the more likely they are to
be perceived as causally related (e.g., Dickinson, Shanks,
& Evenden, 1984; Michotte, 1963). People also tend to
assume that causes and effects should be of similar mag-
nitude. Einhorn and Hogarth (1986) pointed out that the
germ theory of disease initially met with great resistance
because people could not imagine how such tiny organ-
isms could have such devastating effects.
It is important to recognize that homeopathy and
contagion often point us in the right direction. Immu-
nization can be seen as a form of homeopathic medicine
that has an underlying theoretical principle to support it.
My reading of these observations, however, is not that
specific theoretical (causal) principles are constraining
similarity but rather that similarity (homeopathy and
contagion) acts as a constraint on the search for causal
explanations. Even in classical conditioning studies, the
similarity of the conditioned stimulus and the uncondi-
tioned stimulus can have a major influence on the rate
of conditioning (Testa, 1974). Of course, similarity must
itself be constrained for terms like homeopathy to have
a meaning. Shortly, I will suggest some constraints on
similarity as part of an effort to define a role for similarity
in conceptual development.
Similarity is likely to have a significant effect on ex-
planations in another way. Given the importance of sim-
ilarity in retrieval, it is likely that explanations that are
applied to a novel event are constrained by similar events
and their associated explanations. For example, Read
(1983) found that people may rely on single, similar in-
stances in making causal attributions about behaviors.
Furthermore, Ross (1984) and Gentner and Landers
(1985) have found that superficial similarities and not
just similarity with respect to deeper principles or rela-
tions play a major role in determining the remindings
associated with problem solving and the use of analogy.
In brief, it seems that similarity cannot be banished
from the world of theories and conceptual structures. But
it seems to me that a theory of similarity is needed that
is quite different in character from the one summarized
in Table 1. I will suggest an alternative view of similarity
and then attempt to show its value in integrating and
explanation with respect to concepts.
Similarity and Theory in
Conceptual Structure
,4 Contrasting Similarity Model
The following are key tenets of the type of similarity the-
ory needed to link similarity with knowledge-based cat-
egorization: (a) Similarity needs to include attributes, re-
lations, and higher-order relations. (b) Properties in gen-
eral are not independent but rather are linked by a variety
of interproperty relations. (c) Properties exist at multiple
levels of abstraction. (d) Concepts are more than lists.
Properties and relations create depth or structure. Each
of the four main ideas directly conflicts with the corre-
sponding assumption of the theory of similarity outlined
earlier. In one way or another all of these assumptions
are tied to structure. The general idea I am proposing is
far from new. In the psychology of visual perception, the
need for structural approaches to similarity has been a
continuing, if not major, theme (e.g., Biederman, 1985,
1987; Palmer, 1975, 1978; Pomerantz, Sager, & Stoever,
1977). Oden and Lopes (1982) have argued that this view
can inform our understanding of concepts: "Although
similarity must function at some level in the induction
of concepts, the induced categories are not 'held together'
subjectively by the undifferentiated 'force' of similarity,
but rather by structural principles" (p. 78). Noninde-
pendence of properties and simple and higher-order re-
lations add a dimension of depth to categorization. Depth
has clear implications for many of the observations that
seem so problematic for probabilistic view theories. I turn
now to the question of how these modified similarity no-
tions may link up with theory-based categorization.
Psychological Essentialism
Despite the overwhelming evidence against the classical
view, there is something about it that is intuitively com-
pelling. Recently I and my colleagues have begun to take
this observation seriously, not for its metaphysical im-
plications but as a piece of psychological data (Medin&
Ortony, 1989; Medin & Wattenmaker, 1987; Warren-
maker, Nakamura, & Medin, 1988). One might call this
framework "psychological essentialism." The main ideas
are as follows: People act as if things (e.g., objects) have
essences or underlying natures that make them the thing
that they are. Furthermore, the essence constrains or gen-
erates properties that may vary in their centrality. One
of the things that theories do is to embody or provide
causal linkages from deeper properties to more superficial
or surface properties. For example, people in our culture
believe that the categories male and female are genetically
determined, but to pick someone out as male or female
we rely on characteristics such as hair length, height, facial
hair, and clothing that represent a mixture of secondary
sexual characteristics and cultural conventions. Although
1476 December 1989 • American Psychologist
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
these characteristics are more unreliable than genetic ev-
idence, they are far from arbitrary. Not only do they have
some validity in a statistical sense, but also they are tied
to our biological and cultural conceptions of
male and
female.
It is important to note that psychological essentialism
refers not to how the world is but rather to how people
approach the world. Wastebaskets probably have no true
essence, although we may act as if they do. Both social
and psychodiagnostic categories are at least partially cul-
ture specific and may have weak if any metaphysical un-
derpinnings (see also Morey & McNamara, 1987).
If psychological essentialism is bad metaphysics, why
should people act as if things had essences? The reason
is that it may prove to be good epistomology. One could
say that people adopt an
essentialist heuristic,
namely,
the hypothesis that things that look alike tend to share
deeper properties (similarities). Our perceptual and con-
ceptual systems appear to have evolved such that the es-
sentialist heuristic is very often correct (Medin & Wat-
tenmaker, 1987; Shepard, 1984). This is true even for
human artifacts such as cars, computers, and camping
stoves because structure and function tend to be corre-
lated. Surface characteristics that are perceptually obvious
or are readily produced on feature listing tasks may not
so much constitute the core of a concept as point toward
it. This observation suggests that classifying on the basis
of similarity will be relatively effective much of the time,
but that similarity will yield to knowledge of deeper prin-
ciples. Thus, in the work of Gelman and Markman
(1986a) discussed earlier, category membership was more
important than perceptual similarity in determining in-
ductive inferences.
Related Evidence
The contrasting similarity principles presented earlier
coupled with psychological essentialism provide a frame-
work for integrating knowledge-based and similarity-
based categorization. Although it is far short of a formal
theory, the framework provides a useful perspective on
many of the issues under discussion in this article.
1. Nonindependence of features.
Earlier I men-
tioned that classifying on the basis of similarity to a pro-
totype was functionally equivalent to adding up the evi-
dence favoring a classification and applying some criterion
(at least X out of Y features). Recall also that the data
ran strongly against this idea. From the perspective cur-
rently under consideration, however, there ought to be
two ways to produce data consistent with prototype the-
ory. One would be to provide a theory that suggests the
prototype as an ideal or that makes summing of evidence
more natural. For example, suppose that the characteristic
properties for one category were as follows: It is made of
metal, has a regular surface, is of medium size, and is
easy to grasp. For a contrasting category the characteristic
properties were: It is made of rubber, has an irregular
surface, is of small size, and is hard to grasp. The cate-
gories may not seem sensible or coherent but suppose
one adds the information that the objects in one category
could serve as substitutes for a hammer. Given this new
information, it becomes easy to add up the properties of
examples in terms of their utility in supporting ham-
mering. In a series of studies using the above descriptions
and related examples, Wattenmaker, Dewey, Murphy, and
I (1986) found data consistent with prototype theory when
the additional information was supplied, and data incon-
sistent with prototype theory when only characteristic
properties were supplied. Specifically, they found that
linearly separable categories were easier to learn than
nonlinearly separable categories only when an organizing
theme was provided (see also Nakamura, 1985).
One might think that prototypes become important
whenever the categories are meaningful. That is not the
case. When themes are provided that are not compatible
with a summing of evidence, the data are inconsistent
with prototype theories. For instance, suppose that the
examples consisted of descriptions of animals and that
the organizing theme was that one category consisted of
prey and the other of predators. It is a good adaptation
for prey to be armored and to live in trees, but an animal
that is both armored and lives in trees may not be better
adapted than an animal with either characteristic alone.
Being armored and living in trees may be somewhat in-
compatible. Other studies by Wattenmaker et al. using
directly analogous materials failed to find any evidence
that linear separability (and, presumably, summing of
evidence) was important or natural. Only some kinds of
interproperty relations are compatible with a summing
of evidence, and evidence favoring prototypes may be
confined to these cases.
The above studies show that the ease or naturalness
of classification tasks cannot be predicted in terms of
abstract category structures based on distribution of fea-
tures, but rather requires an understanding of the knowl-
edge brought to bear on them, for this knowledge deter-
mines inter-property relationships. So far only a few types
of interproperty relationships have been explored in cat-
egorization, and much is to be gained from the careful
study of further types of relations (e.g., see Barr & Caplan,
1987; Chaflin & Hermann, 1987; Rips & Conrad, 1989;
Winston, Chaftin, & Herman, 1987).
2. Levels of features.
Although experimenters can
often contrive to have the features or properties com-
prising stimulus materials at roughly the same level of
abstractness, in more typical circumstances levels may
vary substantially. This fact has critical implications for
descriptions of category structure (see Barsalou & Bill-
man, 1988). This point may be best represented by an
example from some ongoing research I am conducting
with Glenn Nakamura and Ed Wisniewski. Our stimulus
materials consist of children's drawings of people, a sam-
ple of which is shown in Figure 1. There are two sets of
five drawings, one on the left and one on the right. The
task of the participants in this experiment is to come up
with a rule that could be used to correctly classify both
these drawings and new examples that might be presented
later.
One of our primary aims in this study was to ex-
December 1989 • American Psychologist 1477
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
Figure 1
Children's Drawings of People Used in the
Rule Induction Studies by Nakamura,
Wisniewski, and Medin
amine the effects of different types of knowledge structures
on rule induction. Consequently, some participants were
told that one set was done by farm children and the other
by city children; some were told that one set was done
by creative children and the other by noncreative children;
and still others were told that one set was done by emo-
tionally disturbed children and the other by mentally
healthy children. The exact assignment of drawings was
counterbalanced with respect to the categories such that
half the time the drawings on the left of Figure 1 were
labeled as done by farm children and half the time the
drawings on the right were labeled as having been done
by farm children.
Although we were obviously expecting differences
in the various conditions, in some respects the most strik-
ing result is one that held across conditions. Almost with-
out exception the rules that people gave had properties
at two or three different levels of abstractness. For ex-
ample, one person who was told the drawings on the left
were done by city children gave the following rule: "The
city drawings use more profiles, and are more elaborate.
The clothes are more detailed, showing both pockets and
buttons, and the hair is drawn in. The drawings put less
emphasis on proportion and the legs and torso are off."
Another person who was told the same drawings were
done by farm children wrote: "The children draw what
they see in their normal life. The people have overalls on
and some drawings show body muscles as a result of labor.
The drawings are also more detailed. One can see more
facial details and one drawing has colored the clothes and
another one shows the body under the clothes." As one
can see, the rules typically consist of a general assertion
or assertions coupled with either an operational definition
or examples to illustrate and clarify the assertion. In some
cases these definitions or examples extend across several
levels of abstractness.
One might think that our participants used different
levels of description because there was nothing else for
them to do. That is, there may have been no low-level
perceptual features that would separate the groups. In a
followup study we presented examples one at a time and
asked people to give their rule after each example. If peo-
pie are being forced to use multiple levels of description
because simple rules will not work, then we should ob-
serve a systematic increase in the use of multiple levels
across examples. In fact, however, we observed multiple
levels of description as the predominant strategy from the
first example on. We believe that multiple levels arise
when people try to find a link between abstract explan-
atory principles or ideas (drawings reflect one's experi-
ence) and specific details of drawings.
There are several important consequences of mul-
tilevel descriptions. First of all, the relation across levels
is not necessarily a subset, superset, or a part-whole re-
lation. Most of the time one would say that the lower level
property "supports" the higher level property; for ex-
ample, "jumping into a swimming pool with one's clothes
on" supports poor judgment. This underlines the point
that categorization often involves more than a simple
matching of properties. A related point is that features
are ambiguous in the sense that they may support more
than one higher level property. When the drawings on
the right were associated with the label mentally healthy,
a common description was "all the faces are smiling."
When the label for the same drawing was noncreative, a
common description was "the faces show little variability
in expression." Finally, it should be obvious that whether
a category description is disjunctive (e.g., pig's nose or
cow's mouth or catlike ears) or conjunctive or defining
1478 December 1989 • American Psychologist
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
(e.g., all have animal parts) depends on the level with
respect to which the rule is evaluated.
3. Centrality.
If properties are at different levels of
abstraction and linked by a variety of relations, then one
might imagine that some properties are more central than
others because of the role they play in conceptual struc-
ture. An indication that properties differ in their centrality
comes from a provocative study by Asch and Zukier
(1984). They presented people with trait terms that ap-
peared to be contradictory (e.g., kind and vindictive) and
asked participants if these descriptions could be resolved
(e.g., how could a person be both kind and vindictive?).
Participants had no difficulty integrating the pairs of
terms, and Asch and Zukier identified seven major res-
olution strategies. For present purposes, what is notable
is that many of the resolution strategies involve making
one trait term more central than the other one. For ex-
ample, one way of integrating
kind
and
vindictive
was to
say that the person was fundamentally evil and was kind
only in the service of vindictive ends.
In related work, Shoben and I (Medin & Shoben,
1988) showed that centrality of a property depends on the
concept of which it is a part. We asked participants to
judge the typicality of adjective noun pairs when the ad-
jective was a property that other participants judged was
not true of the noun representing the concept. For example,
our participants judged that all bananas and all boomerangs
are curved. Based on this observation, other participants
were asked to judge the typicality of a straight banana as
a banana or a straight boomerang as a boomerang. Other
instances of the 20 pairs used include
soft knife
versus
soft
diamond
and
polka dot .fire hydrant
versus
polka dot yield
sign.
For 19 of the 20 pairs, participants rated one item
of a pair as more typical than the other. Straight banana,
soft knife, and polka dot fire hydrant were rated as more
typical than straight boomerang, soft diamond, and polka
dot yield sign. In the case of boomerangs (and probably
yield signs), centrality may be driven by structure-function
correlations. Soft diamonds are probably rated as very
atypical because hardness is linked to many other prop-
erties and finding out that diamonds were soft would call
a great deal of other knowledge into question.
Most recently, Woo Kyoung Ahn, Joshua Ruben-
stein, and I have been interviewing clinical psychologists
and psychiatrists concerning their understanding of psy-
chodiagnostic categories. Although our project is not far
enough along to report any detailed results, it is clear that
the
DSM-IIIR
guidebook (American Psychiatric Asso-
ciation, 1987) provides only a skeletal outline that is
brought to life by theories and causal scenarios underlying
and intertwined with the symptoms that comprise the
diagnostic criteria. Symptoms differ in the level of ab-
stractness and the types and number of intersymptom
relations in which they participate, and as a consequence,
they differ in their centrality.
Conclusions
The shift to a focus on knowledge-based categorization
does not mean that the notion of similarity must be left
behind. But we do n~ed an updated approach to, and
interpretation of, similarity. The mounting evidence on
the role of theories and explanations in organizing cate-
gories is much more compatible with features at varying
levels linked by a variety of interproperty relations than
it is with independent features at a single level. In addition,
similarity may not so much constitute structure as point
toward it. There is a dimension of depth to categorization.
The conjectures about psychological essentialism may be
one way of reconciling classification in terms of perceptual
similarity or surface properties with the deeper substance
of knowledge-rich, theory-based categorization.
REFERENCES
Abelson, R. P., & I_alljee, M. G. (1988). Knowledge-structures and causal
explanation. In D. J. Hilton (Ed.),
Contemporary science and natural
explanation: Commonsense conceptions of causality (pp. 175-202).
Brighton, England: Harvester Press.
Allen, S. W., Brooks, L. R., Norman, G. R., & Rosenthal, D. (1988,
November). Effect of prior examples on rule-based diagnostic perfor-
mance. Paper presented at the meeting of the Psychonomic Society,
Chicago.
American Psychiatric Association. (1987). Diagnostic and statistical
manual of mental disorders (rev. ed.). Washington, DC: Author.
Anderson, J. R. (1988). The place of cognitive architectures in a rational
analyses. In The Tenth Annual Conference of the Cognitive Science
Society (pp. 1-10). Montreal, Canada: University of Montreal.
Arkes, H. R., & Harkness, A. R. (1980). Effect of making a diagnosis
on subsequent recognition of symptoms. Journal of Experimental
Psychology: Human Learning and Memory, 6, 568-575.
Armstrong, S. L., Gleitman, L. R., & Gleitman, H. (1983). What some
concepts might not be. Cognition. 13, 263-308.
Asch, S. E., & Zukier, H. (1984). Thinking about persons. Journal of
Personality and Social Psychology,, 46, 1230-1240.
Barr, R. A., & Caplan, L. J. (1987). Category representations and their
implications for category structure. Memory and Cognition, 15, 397-
418.
Barsalou, L. W. (1983). Ad hoc categories. Memory and Cognition, 11,
211-227.
Barsalou, L. W. (1985). Ideals, central tendency, and frequency of in-
stantiation as determinants of graded structure in categories. Journal
of Experimental Psychology: Learning, Memory and Cognition, 11,
629-654.
Barsalou, L. W. (1987). The instability of graded structure: Implications
for the nature of concepts. In U. Neisser (Ed.), Concepts and conceptual
development: The ecological and intellectual factors in categorization
(pp. 101-140). Cambridge, England: Cambridge University Press.
Barsalou, L. W. (1989). Intra-concept similarity and its implications for
inter-concept similarity. In S. Vosniadou & A. Ortony (Eds.), Similarity
and analogical reasoning (pp. 76-121). Cambridge, England: Cam-
bridge University Press.
Barsalou, L. W., & Billman, D. (1988, April). Systematicity and semantic
ambiguity. Paper presented at a workshop on semantic ambiguity,
Adelphi University.
Barsalou, L. W., & Medin, D. L. (1986). Concepts: Fixed definitions or
dynamic context-dependent representations? Cahiers de Psychologic
Cognitive, 6, 187-202.
Bellezza, E S. (1984). Reliability of retrieval from semantic memory:
Noun meanings. Bulletin of the Psychonomic Society, 22, 377-380.
Biederman, I. (1985). Human image understanding: Recent research
and a theory. Computer Fision, Graphics, and Image Processing, 32,
29-83.
Biederman, I. (1987). Recognition-by-components: A theory of human
image understanding.
Psychological Review, 94,
115-147.
Cantor, N., & Mischel, W. (1977). Traits as prototypes: Effects on rec-
ognition memory.
Journal of Personality and Social Psychology, 35,
38-48.
December 1989 • American Psychologist 1479
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
Carey, S. (1985). Conceptual change in childhood. Cambridge, MA:
Massachusetts Institute of Technology Press.
Chattin, R., & Herrmann, D. J. (1987). Relation element theory: A new
account of the representation and processing of semantic relations.
In D. Gorfein & R. Hoffman (Eds.), Learning and memory." The Eb-
binghaus centennial conference (pp. 221-245). Hillsdale, NJ: Erlbaum.
Dickinson, A., Shanks, D., & Evenden, J. (1984). Judgment of act-out-
comes contingency: The role of selective attribution. Quarterly Journal
of Experimental Psychology, 36A(1), 29-50.
Einhorn, J. H., & Hogarth, R. M. (1986). Judging probable cause. Psy-
chological Bulletin, 99, 3-19.
Fates, W. K. (1986a). Memory storage and retrieval processes in category
learning. Journal of Experimental Psychology: General, 115, 155-
175.
Estes, W. K. (1986b). Array models for category learning. Cognitive
Psychology, 18, 500-549.
Fehr, B. (1988). Prototype analysis of the concepts of love and commit-
ment. Journal of Personality and Social Psychology, 55, 557-579.
Fehr, B., & Russell, J. A. (1984). Concept of emotion viewed from a
prototype perspective. Journal of Experimental Psychology." General,
113, 464--486.
Flannagan, M. J., Fried, L. S., & Holyoak, K. J. (1986). Distributional
expectations and the induction of category structure. Journal of Ex-
perimental Psychology: Learning, Memory and Cognition, 12, 241-
256.
Frazer, J. G. (1959). The new golden bough. New York: Criterion Books.
Fried, L. S., & Holyoak, K. J. (1984). Induction of category distribution:
A framework for classification learning. Journal of Experimental Psy-
chology: Learning, Memory and Cognition, 10, 234--257.
Gati, I., & Tversky, A. ( i 984). Weighting common and distinctive features
in perceptual and conceptual judgments. Cognitive Psychology,, 16,
341-370.
Crelman, S. A. (1988). The development of induction within natural
kind and artifact categories. Cognitive Psychology, 20, 65-95.
Gelman, S. A., & Markman, E. M. (1986a). Categories and induction
in young children. Cognition, 23, 183-209.
Gelman, S. A., & Markman, E. M. (1986b). Young children's inductions
from natural kinds: The role of categories and appearances. Child
Development, 58, 1532-1541.
Genero, N., & Cantor, N. (1987). Exemplar prototypes and clinical di-
agnosis: Toward a cognitive economy. Journal of Social and Clinical
Psychology, 5, 59-78.
Gentner, D., & Landers, R. (1985). Analogical reminding: A good match
is hard to find. Paper presented at the International Conference of
Systems, Man and Cybernetics, Tucson, AZ.
Goldin, S. E. (1978). Memory for the ordinary: Typicality effects in
chess memory. Journal of Experimental Psychology: Human Learning
and Memory, 4, 605-616.
Goodman, N. (1972). Seven strictures on similarity. In N. Goodman
(Ed.), Problems and projects. New York: Bobbs-Merrill.
Hartley, J., & Homa, D. (198 I). Abstraction of stylistic concepts. Journal
of Experimental Psychology: Human Learning and Memory, 7, 33-
46.
Hilton, D. J., & Slugoski, B. R. (1986). Knowledge-based causal attri-
bution: The abnormal conditions focus model. Psychological Review,
93, 75-88.
Hintzman, D. L. (1986). "Schema abstraction" in a multiple-trace
memory model. Psychological Review, 93, 411--428.
Homa, D. (1984). On the nature of categories. In G. Bower (Ed.), The
psychology of learning and motivation (Vol. 18, pp. 49-94). New York:
Academic Press.
Homa, D., & Vosburgh, R. (1976). Category breadth and the abstraction
of prototypical information. JEP: Human Learning and Memory, 2,
322-330.
Keil, E C. (1979). Semantic and conceptual development: An ontological
perspective. Cambridge, MA: Harvard University Press.
Keil, E C. ( 1981). Constraints on knowledge and cognitive development.
Psychological Review,, 88, 197-227.
Keil, E C. (1986). The acquisition of natural kind and artifact terms.
In W. Demopoulos & A. Marras (Eds.), Language learning and concept
acquisition (pp. 133-153). Norwood, NJ: Ablex.
Keil, E C. (1987). Conceptual development and category structure. In
U. Neisser (Ed.), Concepts and conceptual development: Ecological
and intellectual factors in categorization (pp. 175-200). Cambridge,
England: Cambridge University Press.
Keil, E C., & Kelly, M. H. (1987). Developmental changes in category
structure. In S. Harnad (Ed.), Categoricalperception: The groundwork
of cognition (pp. 491-510). Cambridge, England: Cambridge Uni-
versity Press.
Kemler-Nelson, D. G. (1984). The effect of intention on what concepts
are acquired. Journal of Verbal Learning and Verbal Behavior, 23,
734-759.
Kolodner, J. L. (1984). Retrieval and organizational structures in con-
ceptual memory: A computer model. Hillsdale, N J: Eflbaum.
Lakoff, G. (1987). Women, fire, and dangerous things: What categories
tell us about the nature of thought. Chicago: University of Chicago
Press.
Leddo, J., Abelson, R. P., & Gross, P. H.(1984). Conjunctive explanation:
When two explanations are better than one. Journal of Personality
and Social Psychology, 47, 933-943.
Malt, B. C., & Smith, E. E. (1983). Correlated properties in natural
categories. Journal of Verbal Learning and Verbal Behavior, 23, 250-
269.
Markman, E. M. (1987). How children constrain the possible meanings
of words. In U. Neisser (Ed.), Concepts and conceptual development:
The ecological and intellectual factors in categorization (pp. 256-
287). Cambridge, England: Cambridge University Press.
Massey, C. M., & Gelman, R. (1988). Preschoolers" ability to decide
whether a photographed unfamiliar object can move itself. Develop-
mental Psychology, 24, 307-317.
McCloskey, M., & Glucksber~ S. (1978). Natural categories: Well-defined
or fuzzy sets? Memory and Cognition, 6, 462--472.
McNamara, T. P., & Sternberg, R. J. (1983). Mental models of word
meaning. Journal of Verbal Learning and Verbal Behavior, 22, 449-
474.
Medin, D. L. (1986). Commentary on "Memory storage and retrieval
processes in category learning." Journal of Experimental Psychology."
General, 115(4), 373-381.
Medin, D. L., Altom, M. W., Edelson, S. M., & Freko, D. (1982). Cor-
related symptoms and simulated medical classification. Journal of
Experimental Psychology: Learning, Memory and Cognition, 8, 37-
50.
Medin, D. L., & Ortony, A. (t989). Psychological essentialism. In S.
Vosniadou & A. Ortony (Eds.), Similarity and analogical reasoning
(pp. 179-195). New York: Cambridge University Press.
Medin, D. L., & Ross, B. H. (1989). The specific character of abstract
thought: Categorization, problem-solving, and induction. In R. J.
Sternberg (Ed.), Advances in the psychology of human intelligence
(Vol. 5, pp. 189-223). Hillsdale, NJ: Erlbaum.
Medin, D. L., & Schaffea; M. M. (1978). A context theory of classification
learning. Psychological Review, 85, 207-238.
Medin, D. L., & Schwanenflugel, P. J. (1981). Linear separability in
classification learning. Journal of Experimental Psychology: Human
Learning and Memory, 7, 355-368.
Medin, D. L., & Shoben, E. J. (1988). Context and structure in conceptual
combination. Cognitive Psychology, 20, 158-190.
Medin, D. L., & Smith, E. E. (1984). Concepts and concept formation.
In M. R. Rosenzweig & L. W. Porter (Eds.), Annual Review of Psy-
chology, 35, 113-118.
Medin, D. L., & Wattenmaker, W. D. (1987). Category cohesiveness,
theories, and cognitive archeology. In U. Neisser (Ed.), Concepts and
conceptual development: The ecological and intellectual factors in cat-
egories (pp. 25-62). Cambridge, England: Cambridge University Press.
Mervis, C. B., & Rosch, E. (1981). Categorization of natural objects. In
M. R. Rosenzweig & L. W. Porter (Eds.), Annual Review of Psychology,
32, 89-115.
Michotte, A. (1963). Perception of causality. London: Methuen.
Morey, L. C., & McNamara, T. P. (1987). On definitions, diagnosis, and
DSM-III. Journal of Abnormal Psychology, 96, 283-285.
Murphy, G. L., & Medin, D. L. (1985). The role of theories in conceptual
coherence. Psychological Review,, 92, 289-316.
Nakamura, G. V. (1985). Knowledge-based classification of ill-defined
categories. Memory and Cognition, 13, 377-384.
1480 December 1989
*
American
Psychologist
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
Nosofsky, R. M. (1987). Attention and learning processes in the iden-
tification and categorization of integral stimuli. Journal of Experi-
mental Psychology: Learning, Memo~ and Cognition, 13, 87-108.
Nosofsky, R. M. (1988a). Exemplar-based accounts of relations between
classification, recognition, and typicality. Journal of Experimental
Psychology: Learning, Memory, and Cognition, 14, 700-708.
Nosofsky, R. M. (1988b). Similarity, frequency, and category represen-
tations. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 14, 54-65.
Oden, G. C. (1987). Concept, knowledge, and thought. In M. R. Ro-
senzweig & L. W. Porter (Eds.), Annual Review of Psychology, 38,
203-227.
Oden, G. C., & Lopes, L. (1982). On the internal structure of fuzzy
subjective categories. In R. R. Yager (Ed.), Recent developments in
fuzzy set and possibility theory (pp. 75-89). Elmsford, NY: Pergamon
Press.
Ortony, A., Vondruska, R. J., Fo~s, M. A., & Jones, L. E. (1985). Salience,
similes, and the asymmetry of similarity. Journal of Memory and
Language, 24, 569-594.
Palmer, S. E. (1975). Visual perception and world knowledge. In D. A.
Norman & D. E. Rumelhart (Eds.), Explorations in cognition (pp.
279-307). San Francisco: W. H. Freeman.
Palmer, S. E. (1978). Structural aspects of visual similarity. Memory and
Cognition, 6, 91-97.
Peterson, M. J., Meaghe~ R. B., Jr., Chair, H., & Gillie, S. (1973). The
abstraction and generalization of dot patterns. Cognitive Psychology,
4, 378-398.
Pomerantz, J. R., Sager, L. C., & Stoever, R. G. (1977). Perception of
wholes and their component parts: Some configural superiority effects.
Journal of Experimental Psychology: Human Perception and Perfor-
mance, 3, 422-435.
Read, S. J. (1983). Once is enough: Causal reasoning from a single in-
stance. Journal of Personality and Social Psychology, 45, 323-334.
Rips, L. (1989). Similarity, typicality, and categorization. In S. Vosniadou
& A. Ortony (Eds.), Similarity and analogical reasoning (pp. 21-59).
New York: Cambridge University Press.
Rips, L. J,, & Conrad, E G. (1989). The folk psychology of mental
activities. Psychological Review, 96, 187-207.
Rosch, E. (1975). Cognitive representations of semantic categories.
Journal of Experimental Psychology: General, 104, 192-233.
Rosch, E., & Mervis, C. B. (1975). Family resemblances: Studies in the
internal structure of categories. Cognitive Psychology, 7, 573-605.
Ross, B. H. (1984). Remindings and their effects in learning a cognitive
skill. Cognitive Psychology, 16, 371-416.
Roth, E. M., & Shoben, E. J. (1983). The effect of context on the structure
of categories. Cognitive Psychology, 15, 346-378.
Schank, R. C., Collins, G. C., & Hunter, L. E. (1986). Transcending
induction category formation in learning. The Behavioral and Brain
Sciences, 9, 639-686.
Schweder, R. A. (1977). Likeness and likelihood in everyday thought:
Magical thinking in judgments about personality. Current Anthro-
pology, 18, 4.
Sebestyn, G. S. (1962). Decision-making processes in pattern recognition.
New York: Macmillan.
Shepard, R. H. (1984). Ecological constraints on internal representation:
Resonant kinematics of perceiving, imagining~ thinking, and dreaming.
Psychological Review, 19, 417-447.
Shepard, R. N., Hovland, C. I., & Jenkins, H. M. (1961). Learning and
memorization of classifications. Psychological Monographs, 75, (13,
Whole No. 517).
Smith, E. E., &Medin, D. L. (1981 ). Categories and concepts. Cambridge,
MA: Harvard University Press.
Smith, E. E., Rips, J. J., & Me&n, D. W. (1984). A psychological approach
to concepts: Comments on Rey's "Concepts and stereotypes." Cog-
nition,
17,
265-274.
Smith, E. E., Shoben, E. J., & Rips, J. J. (1974). Structure and processes
in semantic memory: A featural model for semantic decisions. Psy-
chological Review, 81, 214-241.
Strainers, E (1971). Structural ontology. Philosophia, 1, 21-42.
Testa, T. J. (1974). Causal relationships and the acquisition of avoidance
responses. Psychological Review, 81, 491-505.
Tulving, E. (1983). Elements of episodic memory. New York: Oxford
University Press.
Tversky, A. (1977). Features of similarity. Psychological Review, 84, 327-
352.
Watanabe, S. (1969). Knowing and guessing: A formal and quantitative
study. New York: Wiley.
Wattenmaker, W. D., Dewey, G. I., Murphy, T. D., & Merlin, D. L.
(1986). Linear separability and concept learning: Context, relational
properties, and concept naturalness. Cognitive Psychology, 18, 158-
194.
Wattenmaker, W. D., Nakamura, G. V., & Medin, D. L. (1988). Rela-
tionships between similarity-based and explanation-based categori-
zation. In D. Hilton (Ed.), Contemporary science and natural expla-
nation: Commonsense conceptions of causality (pp. 205-241). Brigh-
ton, England: Harvester Press.
Winston, M. E., Chafiin, R., & Herrmann, D. (1987). A taxonomy of
part-whole relations. Cognitive Science, I 1, 417--444.
Wittgenstein, L. (1953). Philosophical investigations ((3. E. M. Ans-
combe, trans.). Oxford, England: Blackwell.
December 1989 * American Psychologist 1481
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.